AI Crypto Trading Bot: 5 Platforms Checked, 1 Real Model (2026)

An AI crypto trading bot is supposed to be software that learns from market data and decides what to trade. That is the promise on every landing page in the category. So we did the boring thing nobody selling one does: we went to five of the biggest crypto bot platforms and read what their own documentation says about AI. Not their homepages — their help centres, their feature indexes, their sitemaps.

Four of the five document no machine learning at all. The one tool in the category with a real, inspectable, peer-reviewed model is free, open source, and its own docs tell you not to run the example in production. This is what we found, how to run the same check yourself in ten minutes, and what two US regulators have already said about the gap between the marketing and the code.

An AI crypto trading bot is a program that places buy and sell orders on a crypto market using a statistical model trained on historical data, rather than a fixed set of rules written by a human. That is the only definition that distinguishes it from the bots that have existed since 2017. A grid bot that buys every 2% down is automation. A model that was trained on 10,000 features, scored on data it never saw, and retrained weekly is machine learning. The word "AI" in this category is doing the work of hiding which one you are buying.

In practice, the label gets attached to four very different things, and only the last one is what the search term implies:

• A rules engine with a marketing refresh. DCA, grid, arbitrage and signal bots, unchanged, with "AI-powered" added to the pricing page. This is the overwhelming majority of the category.
• A chat assistant bolted onto the docs. A large language model that helps you configure the bot or write a strategy. It does not place trades or choose assets.
• An MCP server. An API that lets your own AI assistant read your account and your positions. The direction here is backwards from what people expect: the model is outside the bot, querying it, not inside it deciding.
• An actual trained model. Features, labels, a training window, a test set, and inference on unseen data. Rare, and almost always open source.

If you have read our breakdown of the eight types of trading bot for crypto, the taxonomy will be familiar. What has changed in 2026 is that the AI label now floats freely across all eight categories, which makes it useless as a filter. You need a test instead.

Machine learning has a fixed vocabulary, and any product genuinely using it has to publish five specific things. Ask for them, and the category sorts itself in about ten minutes. We did not invent this list — it is lifted directly from the glossary that the one open-source project in this space puts in its own introduction, before any feature claims.

• Features. What inputs does the model see? These are the parameters, derived from historical data, that the model is trained on. A real product can tell you the family of inputs even if it will not list all of them.
• Labels. What is it trained to predict? A label is the target value each input vector is paired with, and labels by definition look forward in time — the price 100 candles ahead, or whether price rose. If a vendor cannot say what the model predicts, there is no model.
• Training. Over what window, and how often does it retrain? A model trained once in 2023 and never updated is a fossil in a market that reorganises every quarter.
• Test data. What was held back? Test data is the slice used to score the model after training, which by construction never influenced its weights. Without a held-out set, a backtest is a memory exercise.
• Inferencing. How does a trained model get fed new, unseen data in live conditions, and how stale is it when it decides?

Note what is not on the list: win rate, number of exchanges, Trustpilot score, or how many users the platform claims. Those are the five numbers vendors lead with and none of them tell you whether a model exists. Ask for features and labels instead. The answers, or the silence, are the whole review.

We pulled the public page index for five established crypto bot platforms and searched every URL for AI-related product documentation. Across 4,290 indexed pages there are three AI pages, and not one of them is a trading model. This is a reproducible check, not an opinion — the method is in the evaluation section below, and the numbers are as of 1 October 2026.

Platform	Public pages indexed	AI product pages	What the AI actually is
3Commas (help centre)	69 articles	1	An MCP server so your assistant can query 3Commas
Cryptohopper	3,120 URLs	1	A paper-trading tournament between four language models
Gunbot	599 URLs	1 (plus one comparison page)	A custom GPT that reads the documentation
Bitsgap	36 URLs	0	No AI page at all
Coinrule	46 URLs	0	No AI page at all

3Commas: 69 help articles, zero about AI trading

The 3Commas help centre is the complete product manual, and it is 69 articles long. Reading the URL slugs alone tells you what the product is: how the DCA bot works, how the grid bot works, the signal bot JSON guide, how to set up a bot with a TradingView indicator start condition, how to build entry signals with multiple built-in indicators. Every one of those is a deterministic rule you configure. The single article matching "AI" is about connecting your AI assistant to the 3Commas MCP server — the reversed-direction integration described above.

That matters because 3Commas is ranked second on the "top 7 best AI crypto trading bots" article sitting on page one of Google for this keyword. The platform itself never makes that claim in its documentation. The claim is in a press release. Our own 3Commas review and the 3Commas alternatives comparison both describe it as a rules-based bot platform, because that is what the manual describes.

Gunbot: the AI is a custom GPT, and Gunbot says so plainly

Gunbot has 599 indexed pages and two mention AI. One is a comparison with a competitor. The other introduces Gunbot Guru, which Gunbot describes as a custom GPT with access to all of its documentation, built to help users write JavaScript strategies and debug AutoConfig jobs. Gunbot is refreshingly direct about the limits: it says the assistant "is not a human support replacement" and that "as with any AI, it is experimental, and we cannot guarantee that it always provides correct information."

Read that against the product. Gunbot Guru helps you write the rules. You still write the rules. The strategy is a hand-coded JavaScript file checking a bid against a lower Bollinger band and a fast moving average against an EMA — exactly the kind of logic the Gunbot review walks through. An assistant that explains indicators is a documentation feature, and a good one. It is not a model trading your account.

Bitsgap and Coinrule: no AI claim in the index at all

Neither Bitsgap nor Coinrule has a single AI page in its public sitemap. Coinrule is the honest one in the whole category almost by accident: the product name says "rule", the interface is if-this-then-that, and the company never pretends otherwise. It still shows up on page one for "ai trading bot free" — ranked there by Google and by listicles, not by its own claim.

Cryptohopper has 3,120 indexed pages and exactly one AI product page, and it is the most interesting document in this whole category — because it is an honest experiment that quietly undercuts the pitch. The AI Trading Arena runs four frontier language models against each other: Gemini 3 Flash from Google, Llama 4 Maverick from Meta, DeepSeek V3.2, and MiMo v2 Pro from Xiaomi. Each starts with $10,000 in USDT, each makes independent trading decisions every five minutes, and each trades the same five pairs — BTC, ETH, SOL, XRP and DOGE — on identical live price data.

Three details in Cryptohopper's own FAQ are worth more than any vendor comparison table:

• It is not real money. Cryptohopper states that the arena "runs in a paper trading environment" and that "no actual cryptocurrency is bought or sold." Portfolio values are tracked as if trades executed. The company gives the reason itself: to benchmark AI trading performance without market impact.
• There is a buy-and-hold benchmark drawn on the chart. The HODL line shows what the same $10,000 would have done sitting in Bitcoin. Cryptohopper says it plainly: if an agent line is above HODL, it is beating buy-and-hold. The leading bot platform in the world built an AI showcase and put the null hypothesis on the same axis.
• The answer to "can I use AI trading strategies here" does not name a model. Asked that question in its own FAQ, Cryptohopper answers that you can build automated strategies, use copy trading to follow top performers, or set up DCA bots. Those are three rules-based products.

The arena is also marked up in the page source as an event, not as a software feature — which is the correct classification, and a more candid one than the marketing copy around it. For what the paid product actually does, our Cryptohopper review covers the strategy designer, the copy bot and the tier limits, none of which involve a trained model.

We are not mocking this. A public, benchmarked, paper-traded bake-off with a buy-and-hold control is better science than anything else in the category. It is just not a product you can buy, and the thing it is testing is whether a general-purpose language model prompted every five minutes can beat holding Bitcoin.

FreqAI, a module inside the open-source Freqtrade project, is the only tool we found in this category that documents a real model end to end — and it is free, not-for-profit, and peer-reviewed. Its own introduction describes it as software for "training a predictive machine learning model to generate market forecasts given a set of input signals", and as "a sandbox for easily deploying robust machine learning libraries on real-time data."

It passes the five-question test without being asked:

• Features. You build them yourself in a strategy file; the docs are built for feature sets in the 10,000-plus range, with outlier detection and principal component analysis available to cut the dimensionality back down.
• Labels. You define the target. The docs give both shapes explicitly: a regressor predicts a continuous value such as the price 100 candles ahead, a classifier predicts a discrete one such as up or down. The documentation states that labels "intentionally look into the future."
• Training. The example configuration ships with a 30-day training window and retrains on a rolling basis. In live mode it can retrain continuously on a separate thread or GPU so the model self-adapts.
• Test data. The glossary defines train and test splits separately and states that test data "does not influence nodal weights within the model."
• Inferencing. Defined, named, and threaded separately from training so predictions stay fast while models refresh in the background.

The model libraries are named and swappable: LightGBM and XGBoost for regression, classification and multi-target problems, a PyTorch module for multilayer perceptrons and custom neural networks, and a reinforcement-learning build. CatBoost is still documented but the docs note it is no longer actively supported as of version 2025.12 — which is itself a signal of maintenance you will not get from a closed product.

And it is cited. FreqAI is published in the Journal of Open Source Software as "FreqAI: generalizing adaptive modeling for chaotic time-series market forecasts" (2022, DOI 10.21105/joss.04864), with ten named authors. That is the only peer-reviewed citation we found anywhere in this keyword space. If you want to understand how one of these systems is actually assembled, the architecture overlaps heavily with what we documented in building a copy trading bot in Python: the hard parts are data plumbing and execution, not the model.

One more detail says everything about the state of the category. FreqAI puts this note at the top of its introduction: it is and always will be a not-for-profit open source project, it does not have a crypto token, it does not sell signals, and it has no domain other than the Freqtrade documentation. A project has to write that sentence because people are impersonating it to sell things.

The most valuable pages in this entire research pass were the warnings in the only real machine-learning documentation, because each one names a cost that AI bot vendors are silent about. These are not criticisms of FreqAI. They are the honest operating constraints of running a model against a market, written down by the people who built one.

• The example strategy is explicitly not for production. The docs carry a danger block saying the bundled strategy exists to showcase features and run on small computers as a developer benchmark, and is "not designed to be run in production." The thing you can start in one command is a demo.
• You cannot shortcut validation. A true backtest of adaptive training retrains the model once per backtest window, so the docs state that fully testing a model is best done by running it dry and letting it train constantly — and that in that case "backtesting would take the exact same amount of time as a dry run." Thirty days of evidence costs thirty days. There is no fast-forward button, which is precisely what every "backtested 500% returns" screenshot is implying there is.
• Models go stale while they wait their turn. Training is sequential per pair. The docs work the arithmetic: 50 pairs at five minutes each means the oldest model is over four hours old. There is a configuration setting to refuse entries from models beyond an age limit, which exists because this is a real problem. If your target trade duration is shorter than your model age, you are trading on a stale forecast.
• Look-ahead bias is a documented trap, not a hypothetical. Features are computed once over the whole training range, so the docs warn you to be sure your features do not look into the future. This is the single most common way a backtest produces a number that cannot be repeated with real money.
• Downloading somebody else's model file is a security risk. Loading saved models requires an unsafe deserialisation flag, and the docs state the risk is absent only as long as you load models you trained yourself. Any marketplace selling you a pre-trained "AI strategy" is asking you to execute a stranger's file.

Put those five together and you get the real cost of a genuine AI trading bot: weeks of dry-running before you learn anything, a model that is hours old at decision time, a feature pipeline you have to audit for leakage yourself, and no safe way to outsource any of it. That is why the honest version is free and the expensive versions are rules engines. The same uncomfortable arithmetic we ran in the math before the charts applies here: automation changes who clicks the button, not whether the edge exists.

Yes, but almost nothing ranking for "free AI trading bot" is both free and AI. We pulled the Google results for that query the same day as the main keyword. Page one was two app-store listings, a forex platform product page, a bot vendor tool page, a comparison site, and a press release titled around eight free automated AI trading bots for beginners. The genuinely free, genuinely model-driven option — FreqAI — was not on it.

The main keyword SERP is no better. Alongside a Reddit thread from a beginner stuck on backtesting and one decent explainer, page one holds an App Store listing, a press release ranking seven products, a Fiverr category page selling freelance bot-building gigs, and a listicle that calls an exchange grid bot a "free AI crypto trading bot" — a grid bot being the single most deterministic strategy in crypto. Google is surfacing advertising because advertising is most of what has been published on the term.

Free in this category usually means one of four things, and it is worth knowing which you are being offered:

• Free tier, paid limits. The bot runs but caps positions, pairs, or how often it checks the market. Both the free and paid tiers run the same rules.
• Free to install, you pay the infrastructure. Self-hosted and open-source bots cost nothing to license and then cost a server, RPC access, and your time. FreqAI sits here, and the GPU is not optional if you want fast retraining.
• Free because the exchange earns the spread. Exchange-native bots are bundled with trading fees. You pay on every fill instead of every month, which on an active strategy is usually more.
• Free because you are the product. An app that wants a deposit, or a signal group that wants a referral code. This is the one the regulators wrote about.

We will name our own pricing since we are in this market: uwuu charges a performance-based fee, so you pay when a copied trade profits and nothing when it does not. It is not free, and we do not run a free trial. We think that is the right structure for the same reason the arithmetic above matters — a subscription charges you whether or not the strategy worked.

Two US regulators have already acted on this exact gap between AI marketing and AI code, and both documents are free to read. If you are evaluating a product in this category, these are more useful than any review.

The SEC has a name for it: AI washing

On 18 March 2024 the SEC announced settled charges against two investment advisers, Delphia (USA) Inc. and Global Predictions Inc., for false and misleading statements about their use of artificial intelligence. They paid $400,000 in civil penalties between them — $225,000 and $175,000 respectively. The SEC found that Delphia had claimed, from at least August 2019 to August 2023, that it put collective client data to work to make its artificial intelligence smarter so it could predict which companies and trends were about to make it big. It did not have those capabilities. Global Predictions called itself the "first regulated AI financial advisor" and advertised expert AI-driven forecasts, and the SEC found it did neither.

Both were charged under the antifraud provisions of the Advisers Act and under the Marketing Rule, which prohibits a registered adviser from running an advertisement containing an untrue statement of material fact. The agency describes the practice as "AI washing", by analogy to greenwashing. The enforcement point generalises well beyond registered advisers: claiming a model you do not have is a misstatement, not a style choice.

The CFTC wrote the headline for this article first

On 25 January 2024, the CFTC Office of Customer Education and Outreach published a customer advisory titled, in full, "AI Won't Turn Trading Bots into Money Machines". It opens by saying fraudsters are exploiting public interest in AI to promote automated trading algorithms, trade signal strategies and crypto-asset schemes promising unreasonably high or guaranteed returns, and states directly that AI technology cannot predict the future or sudden market changes. It notes the specific shapes the claims take: returns in the tens of thousands of percent, or 100 percent win rates.

Its case study is the largest of its kind. In the Mirror Trading International matter, customers lost nearly 30,000 bitcoins — worth about $1.7 billion at the time — to a scheme marketed on the strength of an automated trading bot. The advisory's checklist before you fund anything is short and genuinely useful: research the background of the company and the people, including a reverse image search on key personnel; check how old the domain is; get a second opinion from someone you trust; and work out what fees, spreads and subscription costs do to your returns.

That last item is the one traders skip. It is the same point as our breakdown of copy trading risk management: the cost stack decides the outcome far more often than the strategy does.

The useful question is not whether a bot uses AI. It is whether you can audit the thing making the decisions. Those are different questions with very different answers, and this is where we have an obvious interest, so here is the comparison stated against ourselves as well as for us.

Dimension	Marketed AI bot	Documented ML bot (FreqAI)	On-chain copy trading
Who decides the trade	Undisclosed rules	A model you trained	A named human wallet
Can you verify the record	No — vendor screenshots	Yes, on your own dry run	Yes, every trade is on-chain
Setup effort	Minutes	Weeks, plus a GPU	Minutes
What you pay	Subscription, win or lose	Infrastructure and time	Performance fee on profit
Main failure mode	There was never an edge	Overfitting and stale models	Picking the wrong trader

To be explicit, because this article is about not overclaiming: uwuu is not an AI trading bot and we do not describe it as one. There is no predictive model and nothing is forecasting price. It mirrors transactions from a wallet you selected, non-custodially, with sub-400ms execution and rules-based trade filtering you configure. The decision-maker is a human being whose entire history is public on Solana, which is why our ranked list of Solana traders publishes addresses, 30-day PnL and win rates rather than a performance claim you have to take on faith.

That is a narrower promise than "AI picks winners for you", and it is the one we can evidence. It also has a real failure mode, and it is the obvious one: copying a trader who stops working. Our analysis of whether copy trading is profitable is blunt about the distribution, and the answer there is trader selection, not a model. If you want the mechanics first, start with what crypto copy trading is or the Solana trading bot guide.

Everything in this article came from public pages and took one research session. You can run the same audit on any product in the category before you connect an API key. In order of how much time each step saves you:

• Read the sitemap, not the homepage. Fetch the platform sitemap or help-centre index and search the URLs for "ai", "model", "machine-learning" and "neural". A product built around a model has documentation about the model. The slugs are a product map that marketing does not curate.
• Search for the absence. If "features", "labels", "training" and "inference" appear nowhere in the docs of a product sold as AI, you have your answer, and it took two minutes. This is how we established that four of five platforms document rules, not models.
• Check which direction the integration runs. An MCP server or an API for your assistant is a read interface, not a trading brain. Useful, but it does not decide anything.
• Ask what the benchmark is. Returns without a comparison are decoration. Cryptohopper drew a buy-and-hold line on its own AI chart; if a vendor will not name the benchmark its model beat, assume the benchmark is the problem.
• Price the full stack. Subscription, exchange or swap fees, spread, slippage and infrastructure. The CFTC advisory lists this and it is the step that eliminates most strategies on arithmetic alone, before any question of edge.
• Check the publication date and whether it is paid placement. If a product ranks through a press release or a freelance marketplace rather than its own documentation, you are reading an advertisement with a Google ranking.

Run that on the next product you are considering. If you want the same treatment applied to adjacent categories, we have done it for memecoin trading bots, Telegram trading bots and analytics terminals like Sharpe AI, where the AI question comes up for the same reasons.

Can you make money with AI trading bots?

Some people do, but not because of the AI label. The CFTC states directly that AI cannot predict the future or sudden market changes, and the only tool we found with a documented model is an open-source research framework whose own docs say the fastest way to validate it is to run it live for as long as you want evidence. Any product promising reliable returns from AI is making a claim the SEC has already fined firms for.

Do crypto trading bots make money?

A bot is an execution tool, not an edge. It removes hesitation and trades while you sleep, and it will execute a losing strategy with the same discipline as a winning one. Across the categories we have tested, the deciding factors are the fee stack and what the strategy is actually based on, not the automation itself.

Can ChatGPT build a trading bot?

It can write the code, and vendors already use it for exactly that. Gunbot ships a custom GPT that helps users write JavaScript strategies and configuration jobs, while telling users it is experimental and may be wrong. Writing the code is the easy part. Knowing what the strategy should be, and validating it without look-ahead bias, is the part no assistant solves.

Which AI bot is best for trading for beginners?

For a beginner, the honest answer is that no model-driven bot in this category is beginner-appropriate — the real ones need a training pipeline, a held-out test set and weeks of dry running. If the appeal is not wanting to pick trades yourself, copying a trader whose full history is publicly verifiable on-chain is a far more auditable starting point than a model whose inputs nobody will show you.

Is there a free AI crypto trading bot?

FreqAI is genuinely free, open source and not-for-profit, and its maintainers state that it has no token and does not sell signals. It is also the most technically demanding option here. Most products marketed as free AI bots are either free tiers of rules-based bots, exchange bots paid for through trading fees, or apps that need a deposit first.

How do I tell if a trading bot really uses AI?

Ask for five things: the features, the labels, the training window and retrain frequency, the held-out test data, and how inference runs live. Every genuine machine-learning product publishes that vocabulary because it cannot be built without it. If a vendor answers with win rates and user counts instead, it is a rules engine with an AI sticker.

Can you make $1,000 a day with crypto?

Not reliably, and not because of a bot. Making $1,000 a day consistently requires both capital and a repeatable edge, and the maths on position size and cost per round trip is unforgiving at small account sizes. Claims of guaranteed daily returns are the exact pattern the CFTC advisory was written to warn about.